A Novel Method for Road Anomaly Objects Detection in the Traffic Environment With Multi-Mechanism Fusion

In the modern automotive industry, Advanced Driving Assistance Systems (ADAS) have gradually become a standard feature in various types of vehicles, with the important function of detecting road anomalies. The appearance of anomalies on the road can be attributed to unexpected situations while driving, and the current methods for detecting distant or small anomalies are not highly accurate. Therefore, in this paper, a method is proposed that uses semantic segmentation to extract key features from the image, and obtaining a new synthesized image by image resynthesis. Then, segmentation uncertainty and depth information are used to compare the differences between multiple feature maps and the input image to highlight the anomalies. Additionally, a postprocessor is designed to use an anomaly score to enhance the recognition of anomaly target and reduce false positives caused by noise. Experiments are conducted on the Obstacle Track dataset and the Lost and Found dataset, and various methods for detecting anomaly objects are compared. The experimental results demonstrate that the method proposed in this paper can effectively detect un-common objects in the training dataset in road anomaly object detection. It improves the detection rate and reduces the false positive rate based on previous anomaly detection methods. The proposed method presented in this paper achieves high detection rates for both seen and unseen anomaly objects in the training set, which enhances the generalization ability of anomaly detection in the road area of interest.